This paper presents EmoVoice AI, a deep learning-based speech emotion recognition system that classifies human emotional states directly from raw audio signals. The proposed framework adopts a hybrid Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) architecture, jointly capturing localized spectral patterns and long-range temporal dependencies. Acoustic features are derived using Librosa as forty Mel-Frequency Cepstral Coefficients (MFCCs) augmented with first-and second-order delta coefficients, yielding a 120-dimensional representation per frame across a fixed window of 124 frames. The combined IES and SAVEE corpora contribute 955 labelled utterances spanning six emotion classes: anger, disgust, fear, happiness, neutral and sadness. The CNN front end stacks three one-dimensional convolutional blocks with batch normalization, ReLU activation and max pooling, while the temporal back end employs a two-layer BiLSTM with 128 hidden units per direction. A fully connected classification head with dropout regularization produces the softmax output. The model is optimized with Adam and weighted cross-entropy loss to mitigate class imbalance. Under a stratified 80/20 train-test split, the system achieves a training accuracy of 98.82 % and a testing accuracy of 83.25 %, indicating strong generalization on unseen samples. The complete pipeline is implemented in PyTorch and is reproducible in a standard Jupyter Notebook environment, establishing a lightweight, end-to-end baseline suitable for affective computing applications such as mental-health monitoring, intelligent assistants and emotion-aware humancomputer interaction.
Charan et al. (Thu,) studied this question.